After child subagents complete, the coalesced announcement runs as an
internal turn which rolls back all messages from the parent's in-memory
context. This causes the parent LLM to lose findings in subsequent turns.
Add persistResponse option to writeInternal that re-injects the LLM's
summary as a non-internal assistant message after the internal run
completes. The internal prompt stays hidden while the summary persists
in both memory and session JSONL for future turns.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Announcement messages from subagent completion flows were persisted as
regular user messages, polluting conversation history and leaking
orchestration instructions to frontend/CLI.
- Add `internal?: boolean` to SessionEntry message variant
- SessionManager.saveMessage accepts { internal: true } option
- SessionManager.loadMessages filters internal by default
- Agent.runInternal() tags messages as internal, rolls back from memory
- Agent.withRunMutex() prevents concurrent run/runInternal mis-tagging
- AsyncAgent.writeInternal() suppresses event forwarding during internal runs
- Announce flows use writeInternal() instead of write()
- Desktop IPC getHistory reads from session storage (filtered)
- CLI session show parses SessionEntry, supports --show-internal flag
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Cover formatCoalescedAnnouncementMessage (single/multi record,
mixed outcomes, missing findings), registry coalescing state
transitions, shutdown capture, and new field serialization.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When multiple children complete, buffer findings per-child and only
send one combined announcement to the parent after all unannounced
runs for the same requester have finished. This avoids N separate
LLM calls and gives the parent a complete picture of all results.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Resolve tools before building subagent system prompt so the
"## Tooling" section is included, matching OpenClaw's pattern.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
ProfileManager.buildSystemPrompt() now delegates to the structured builder.
Runner assembles prompt after tool resolution with safety, tooling summary,
and runtime info. Subagent prompts use minimal mode with safety constitution.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace for-await stream consumption in watchChildAgent with
waitForIdle() + onClose() callbacks on AsyncAgent. This prevents
conflict with Hub.consumeAgent() which also reads the Channel.
Add shutdownSubagentRegistry() for clean Hub shutdown, guard
Hub access with isHubInitialized(), and clean resumedRuns in sweep.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implement subagent registry, persistence store, and announcement flow
for child agent lifecycle management. Includes types, registry with
auto-archive sweeper, session JSONL reading for result extraction,
and formatted announcement message delivery.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
